Probabilistic genotyping

When a mixture can't be read by eye, software models every possible combination of contributors and reports a number — but the number is only as good as the model.

From eyeballing peaks to running a simulation

For years, examiners interpreted mixed DNA profiles by inspecting the peaks on an electropherogram and deciding, by trained eye, which alleles belonged to which contributor. The method worked tolerably well for clean two-person mixtures with a clear major and minor component. It stopped working reliably when profiles grew complex — three or more contributors, low-template samples where peaks dropped in and out, or near-equal mixtures where no single contributor stood out. The subjectivity was real, and documented: give the same mixture to different examiners and you could get different answers. Studies commissioned by the President's Council of Advisors on Science and Technology ↗ found that some mixture interpretations lacked the foundational validity studies needed to support courtroom use.

Probabilistic genotyping — PG — shifted the task from the examiner's eye to a statistical model. Instead of asking "can I see this contributor's alleles?", the software asks: given all the peaks in this profile, across all their heights and positions, how probable is it that this particular person contributed, compared with the probability that a randomly chosen unrelated person did? The answer is a likelihood ratio: a single number expressing the relative support that the evidence gives to each hypothesis. A ratio of one million means the profile is a million times more probable if the named person contributed than if a random alternative did. A ratio of one means the evidence is neutral.

CROSS-REF

What the model is actually doing

The software — systems in active use include STRmix, TrueAllele, ArmedXpert and others — encodes a probabilistic model of the entire process by which a mixture is generated and measured. It accounts for the expected height of peaks given a notional contributor's genotype, the probability that a low-quantity allele drops out entirely, the probability of drop-in from contamination, the characteristic stutter peaks that PCR amplification always produces alongside genuine signal, and the number of contributors assumed to be present. It then explores — by Markov chain Monte Carlo sampling or similar methods — the enormous space of possible genotype combinations that could have produced the observed pattern, and calculates the likelihood ratio by integrating over that space.

The assumptions baked into that model are where scrutiny belongs. The assumed number of contributors matters: assume two when there are three, and the output is wrong in ways the ratio will not disclose. The parameters used for drop-out rates, stutter ratios and peak height variability are estimated from population data and validation studies; SWGDAM's guidelines for probabilistic genotyping ↗ set expectations for how those validations should be conducted and documented. If a laboratory's population parameters are poorly estimated or their validation set is too small, the reported ratio may be systematically over- or understated — and there is no internal flag to warn a jury of that.

Transparency is a live controversy. Several PG systems are proprietary, and their source code has not been publicly released for independent audit. Defence challenges have sought access to that code; courts in different jurisdictions have reached different conclusions about what disclosure is required. The National Institute of Standards and Technology maintains reference profiles and exercises that allow laboratories to test software performance against known ground truth, but mandatory independent validation before courtroom deployment is not uniformly required.

What probabilistic genotyping genuinely achieves is a disciplined replacement for ad hoc human judgment in situations where that judgment is demonstrably unreliable. What it does not achieve is objectivity independent of its inputs. The model encodes choices — about contributor number, population allele frequencies, drop-out rates — and those choices move the subjectivity upstream, into parameters that are harder for a non-specialist to interrogate than a peak on a chart. The number that emerges is real arithmetic. The question worth asking is what assumptions it was fed.